Multi-scale U-net with Edge Guidance for Multimodal Retinal Image Deformable Registration.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Multimodal retinal image registration is crucial for diagnosing eye diseases like those requiring color fundus and optical coherence tomography (OCT) image alignment.
• Challenges include the difficulty in obtaining ground truth data and limitations of existing rigid registration algorithms that ignore optical distortion.

Purpose of the Study:

• To develop an unsupervised learning method for deformable registration of multimodal retinal images, specifically addressing optical distortion.
• To improve the accuracy and robustness of retinal image registration for clinical applications.

Main Methods:

• Proposed a novel unsupervised learning framework for deformable registration of color fundus and OCT images.
• Incorporated a multi-level receptive field, contour, and local detail considerations.
• Introduced an edge similarity (ES) loss term to address optical distortion differences.
• Utilized a U-net architecture with dilated convolutions, squeeze-and-excitation (SE) blocks, and spatial transformer layers, alongside a multi-scale input layer.

Main Results:

• The proposed framework demonstrated superior performance compared to conventional and deep learning-based methods in quantitative experiments.
• The combination of the ES loss, U-net, and multi-scale layers achieved competitive registration results for both normal and abnormal retinal images.
• The method effectively handles optical distortions inherent in multimodal retinal imaging.

Conclusions:

• The developed unsupervised deformable registration method offers a significant advancement for multimodal retinal image analysis.
• The novel ES loss and architectural components contribute to improved accuracy, particularly in the presence of optical distortions.
• This approach holds promise for enhancing the diagnosis and treatment planning of various eye conditions.